The original motivation here was to fix regressions caused by #22414.
However, while working on this, I ended up discussing a language
simplification with Andrew, which changes things a little from how they
worked before #22414.
The main user-facing change here is that any reference to a prior
function parameter, even if potentially comptime-known at the usage
site or even not analyzed, now makes a function generic. This applies
even if the parameter being referenced is not a `comptime` parameter,
since it could still be populated when performing an inline call. This
is a breaking language change.
The detection of this is done in AstGen; when evaluating a parameter
type or return type, we track whether it referenced any prior parameter,
and if so, we mark this type as being "generic" in ZIR. This will cause
Sema to not evaluate it until the time of instantiation or inline call.
A lovely consequence of this from an implementation perspective is that
it eliminates the need for most of the "generic poison" system. In
particular, `error.GenericPoison` is now completely unnecessary, because
we identify generic expressions earlier in the pipeline; this simplifies
the compiler and avoids redundant work. This also entirely eliminates
the concept of the "generic poison value". The only remnant of this
system is the "generic poison type" (`Type.generic_poison` and
`InternPool.Index.generic_poison_type`). This type is used in two
places:
* During semantic analysis, to represent an unknown result type.
* When storing generic function types, to represent a generic parameter/return type.
It's possible that these use cases should instead use `.none`, but I
leave that investigation to a future adventurer.
One last thing. Prior to #22414, inline calls were a little inefficient,
because they re-evaluated even non-generic parameter types whenever they
were called. Changing this behavior is what ultimately led to #22538.
Well, because the new logic will mark a type expression as generic if
there is any change its resolved type could differ in an inline call,
this redundant work is unnecessary! So, this is another way in which the
new design reduces redundant work and complexity.
Resolves: #22494Resolves: #22532Resolves: #22538
We can still often determine a comptime result based on the type, even
if the pointer is runtime-known.
Also, we previously used load -> is non null instead of AIR
`is_non_null_ptr` if the pointer is comptime-known, but that's a bad
heuristic. Instead, we should check for the pointer to be
comptime-known, *and* for the load to be comptime-known, and only in
that case should we call `Sema.analyzeIsNonNull`.
Resolves: #22556
This was done by regex substitution with `sed`. I then manually went
over the entire diff and fixed any incorrect changes.
This diff also changes a lot of `callconv(.C)` to `callconv(.c)`, since
my regex happened to also trigger here. I opted to leave these changes
in, since they *are* a correct migration, even if they're not the one I
was trying to do!
`Sema.explainWhyValueContainsReferenceToComptimeVar` (concise name!)
adds notes to an error explaining how to get from a given `Value` to a
pointer to some `comptime var` (or a comptime field). Previously, this
error could be very opaque in any case where it wasn't obvious where the
comptime var pointer came from; particularly for type captures. Now, the
error notes explain this to the user.
This rewrite improves some error messages, hugely simplifies the logic,
and fixes several bugs. One of these bugs is technically a new rule
which Andrew and I agreed on: if a parameter has a comptime-only type
but is not declared `comptime`, then the corresponding call argument
should not be *evaluated* at comptime; only resolved. Implementing this
required changing how function types work a little, which in turn
required allowing a new kind of function coercion for some generic use
cases: function coercions are now allowed to implicitly *remove*
`comptime` annotations from parameters with comptime-only types. This is
okay because removing the annotation affects only the call site.
Resolves: #22262
Rather than `Zcu.BuiltinDecl.Memoized` being a struct with fields, it
can instead just be an array, indexed by the enum. This allows runtime
indexing, avoiding a few now-unnecessary `inline` switch cases.
This commit reworks how values like the panic handler function are
memoized during a compiler invocation. Previously, the value was
resolved by whichever analysis requested it first, and cached on `Zcu`.
This is problematic for incremental compilation, as after the initial
resolution, no dependencies are marked by users of this memoized state.
This is arguably acceptable for `std.builtin`, but it's definitely not
acceptable for the panic handler/messages, because those can be set by
the user (`std.builtin.Panic` checks `@import("root").Panic`).
So, here we introduce a new kind of `AnalUnit`, called `memoized_state`.
There are 3 such units:
* `.{ .memoized_state = .va_list }` resolves the type `std.builtin.VaList`
* `.{ .memoized_state = .panic }` resolves `std.Panic`
* `.{ .memoized_state = .main }` resolves everything else we want
These units essentially "bundle" the resolution of their corresponding
declarations, storing the results into fields on `Zcu`. This way, when,
for instance, a function wants to call the panic handler, it simply runs
`ensureMemoizedStateResolved`, registering one dependency, and pulls the
values from the `Zcu`. This "bundling" minimizes dependency edges. The 3
units are separated to allow them to act independently: for instance,
the panic handler can use `std.builtin.Type` without triggering a
dependency loop.
`Zcu.PerThead.ensureTypeUpToDate` is set up in such a way that it only
returns the updated type the first time it is called. In general, that's
okay; however, the exception is that we want the function to continue
returning `error.AnalysisFail` when the type has been lost, or its
number of captures changed.
Therefore, the check for this case now happens before the up-to-date
success return.
For simplicity, the number of captures is now handled by intentionally
losing the instruction in `Zcu.mapOldZirToNew`, since there is nothing
to gain from tracking a type when old instances of it can never be
reused.
The old lowering was kind of neat, but it unintentionally allowed the
syntax `for (123) |_| { ... }`, and there wasn't really a way to fix
that. So, instead, we include both the start and the end of the range in
the `for_len` instruction (each operand to `for` now has *two* entries
in this multi-op instruction). This slightly increases the size of ZIR
for loops of predominantly indexables, but the difference is small
enough that it's not worth complicating ZIR to try and fix it.
Most calls to `requireRuntimeBlock` in Sema are not correct. This
function doesn't deal with all of them, but it does deal with ones which
have, in combination with the past few commits, introduced real-world
regressions.
Related: #22353
This fixes a bug which exposed a compiler implementation detail (ZIR
alloc elision). Previously, `const` declarations with a runtime-known
value in a comptime scope were permitted only if AstGen was able to
elide the alloc in ZIR, since the error was reported by storing to the
comptime alloc.
This just adds a new instruction to also emit this error when the alloc
is elided.
To avoid this PR regressing error messages, most of the work here has
gone towards improving error notes for why code was comptime-evaluated.
ZIR `block_comptime` now stores a "comptime reason", the enum for which
is also used by Sema. There are two types in Sema:
* `ComptimeReason` represents the reason we started evaluating something
at comptime.
* `BlockComptimeReason` represents the reason a given block is evaluated
at comptime; it's either a `ComptimeReason` with an attached source
location, or it's because we're in a function which was called at
comptime (and that function's `Block` should be consulted for the
"parent" reason).
Every `Block` stores a `?BlockComptimeReason`. The old `is_comptime`
field is replaced with a trivial `isComptime()` method which returns
whether that reason is non-`null`.
Lastly, the handling for `block_comptime` has been simplified. It was
previously going through an unnecessary runtime-handling path; now, it
is a trivial sub block exited through a `break_inline` instruction.
Resolves: #22296
This commit separates semantic analysis of the annotated type vs value
of a global declaration, therefore allowing recursive and mutually
recursive values to be declared.
Every `Nav` which undergoes analysis now has *two* corresponding
`AnalUnit`s: `.{ .nav_val = n }` and `.{ .nav_ty = n }`. The `nav_val`
unit is responsible for *fully resolving* the `Nav`: determining its
value, linksection, addrspace, etc. The `nav_ty` unit, on the other
hand, resolves only the information necessary to construct a *pointer*
to the `Nav`: its type, addrspace, etc. (It does also analyze its
linksection, but that could be moved to `nav_val` I think; it doesn't
make any difference).
Analyzing a `nav_ty` for a declaration with no type annotation will just
mark a dependency on the `nav_val`, analyze it, and finish. Conversely,
analyzing a `nav_val` for a declaration *with* a type annotation will
first mark a dependency on the `nav_ty` and analyze it, using this as
the result type when evaluating the value body.
The `nav_val` and `nav_ty` units always have references to one another:
so, if a `Nav`'s type is referenced, its value implicitly is too, and
vice versa. However, these dependencies are trivial, so, to save memory,
are only known implicitly by logic in `resolveReferences`.
In general, analyzing ZIR `decl_val` will only analyze `nav_ty` of the
corresponding `Nav`. There are two exceptions to this. If the
declaration is an `extern` declaration, then we immediately ensure the
`Nav` value is resolved (which doesn't actually require any more
analysis, since such a declaration has no value body anyway).
Additionally, if the resolved type has type tag `.@"fn"`, we again
immediately resolve the `Nav` value. The latter restriction is in place
for two reasons:
* Functions are special, in that their externs are allowed to trivially
alias; i.e. with a declaration `extern fn foo(...)`, you can write
`const bar = foo;`. This is not allowed for non-function externs, and
it means that function types are the only place where it is possible
for a declaration `Nav` to have a `.@"extern"` value without actually
being declared `extern`. We need to identify this situation
immediately so that the `decl_ref` can create a pointer to the *real*
extern `Nav`, not this alias.
* In certain situations, such as taking a pointer to a `Nav`, Sema needs
to queue analysis of a runtime function if the value is a function. To
do this, the function value needs to be known, so we need to resolve
the value immediately upon `&foo` where `foo` is a function.
This restriction is simple to codify into the eventual language
specification, and doesn't limit the utility of this feature in
practice.
A consequence of this commit is that codegen and linking logic needs to
be more careful when looking at `Nav`s. In general:
* When `updateNav` or `updateFunc` is called, it is safe to assume that
the `Nav` being updated (the owner `Nav` for `updateFunc`) is fully
resolved.
* Any `Nav` whose value is/will be an `@"extern"` or a function is fully
resolved; see `Nav.getExtern` for a helper for a common case here.
* Any other `Nav` may only have its type resolved.
This didn't seem to be too tricky to satisfy in any of the existing
codegen/linker backends.
Resolves: #131
The `Cau` abstraction originated from noting that one of the two primary
roles of the legacy `Decl` type was to be the subject of comptime
semantic analysis. However, the data stored in `Cau` has always had some
level of redundancy. While preparing for #131, I went to remove that
redundany, and realised that `Cau` now had exactly one field: `owner`.
This led me to conclude that `Cau` is, in fact, an unnecessary level of
abstraction over what are in reality *fundamentally different* kinds of
analysis unit (`AnalUnit`). Types, `Nav` vals, and `comptime`
declarations are all analyzed in different ways, and trying to treat
them as the same thing is counterproductive!
So, these 3 cases are now different alternatives in `AnalUnit`. To avoid
stealing bits from `InternPool`-based IDs, which are already a little
starved for bits due to the sharding datastructures, `AnalUnit` is
expanded to 64 bits (30 of which are currently unused). This doesn't
impact memory usage too much by default, because we don't store
`AnalUnit`s all too often; however, we do store them a lot under
`-fincremental`, so a non-trivial bump to peak RSS can be observed
there. This will be improved in the future when I made
`InternPool.DepEntry` less memory-inefficient.
`Zcu.PerThread.ensureCauAnalyzed` is split into 3 functions, for each of
the 3 new types of `AnalUnit`. The new logic is much easier to
understand, because it avoids conflating the logic of these
fundamentally different cases.
The new representation is often more compact. It is also more
straightforward to understand: for instance, `extern` is represented on
the `declaration` instruction itself rather than using a special
instruction. The same applies to `var`, making both of these far more
compact.
This commit also separates the type and value bodies of a `declaration`
instruction. This is a prerequisite for #131.
In general, `declaration` now directly encodes details of the syntax
form used, and the embedded ZIR bodies are for actual expressions. The
only exception to this is functions, where ZIR is effectively designed
as if we had #1717. `extern fn` declarations are modeled as
`extern const` with a function type, and normal `fn` definitions are
modeled as `const` with a `func{,_fancy,_inferred}` instruction. This
may change in the future, but improving on this was out of scope for
this commit.
The error messages here aren't amazing yet, but this is an improvement
on status quo, because the current behavior allows false negative
compile errors, so effectively miscompiles.
Resolves: #15874
This code was left over from the legacy Autodoc implementation. No
component of the compiler pipeline actually requires doc comments, so it
is a waste of time and space to store them in ZIR.
Both of these instructions were previously under a special case in
`rvalue` which resulted in every reference to such an instruction adding
a new `ref` instruction. This had the effect that, for instance,
`&a != &a` for parameters. Deduplicating these `ref` instructions was
problematic for different reasons.
For `alloc_inferred`, the problem was that it's not valid to `ref` the
alloc until the allocation has been resolved (`resolve_inferred_alloc`),
but `AstGen.appendBodyWithFixups` would place the `ref` directly after
the `alloc_inferred`. This is solved by bringing
`resolve_inferred_alloc` in line with `make_ptr_const` by having it
*return* the final pointer, rather than modifying `sema.inst_map` of the
original `alloc_inferred`. That way, the `ref` refers to the
`resolve_inferred_alloc` instruction, so is placed immediately after it,
avoiding this issue.
For `param`, the problem is a bit trickier: `param` instructions live in
a body which must contain only `param` instructions, then a
`func{,_inferred,_fancy}`, then a `break_inline`. Moreover, `param`
instructions may be referenced not only by the function body, but also
by other parameters, the return type expression, etc. Each of these
bodies requires separate `ref` instructions. This is solved by pulling
entries out of `ref_table` after evaluating each component of the
function declaration, and appending the refs later on when actually
putting the bodies together. This gives way to another issue: if you
write `fn f(x: T) @TypeOf(x.foo())`, then since `x.foo()` takes a
reference to `x`, this `ref` instruction is now in a comptime context
(outside of the `@TypeOf` ZIR body), so emits a compile error. This is
solved by loosening the rules around `ref` instructions; because they
are not side-effecting, it is okay to allow `ref` of runtime values at
comptime, resulting in a runtime-known value in a comptime scope. We
already apply this mechanism in some cases; for instance, it's why
`runtime_array.len` works in a `comptime` context. In future, we will
want to give similar treatment to many operations in Sema: in general,
it's fine to apply runtime operations at comptime provided they don't
have side effects!
Resolves: #22140
This commit enhances AstGen to introduce a form of error resilience
which allows valid ZIR to be emitted even when AstGen errors occur.
When a non-fatal AstGen error (e.g. `appendErrorNode`) occurs, ZIR
generation is not affected; the error is added to `astgen.errors` and
ultimately to the errors stored in `extra`, but that doesn't stop us
getting valid ZIR. Fatal AstGen errors (e.g. `failNode`) are a bit
trickier. These errors return `error.AnalysisFail`, which is propagated
up the stack. In theory, any parent expression can catch this error and
handle it, continuing ZIR generation whilst throwing away whatever was
lost. For now, we only do this in one place: when creating declarations.
If a call to `fnDecl`, `comptimeDecl`, `globalVarDecl`, etc, returns
`error.AnalysisFail`, the `declaration` instruction is still created,
but its body simply contains the new `extended(astgen_error())`
instruction, which instructs Sema to terminate semantic analysis with a
transitive error. This means that a fatal AstGen error causes the
innermost declaration containing the error to fail, but the rest of the
file remains intact.
If a source file contains parse errors, or an `error.AnalysisFail`
happens when lowering the top-level struct (e.g. there is an error in
one of its fields, or a name has multiple declarations), then lowering
for the entire file fails. Alongside the existing `Zir.hasCompileErrors`
query, this commit introduces `Zir.loweringFailed`, which returns `true`
only in this case.
The end result here is that files with AstGen failures will almost
always still emit valid ZIR, and hence can undergo semantic analysis on
the parts of the file which are (from AstGen's perspective) valid. This
is a noteworthy improvement to UX, but the main motivation here is
actually incremental compilation. Previously, AstGen failures caused
lots of semantic analysis work to be thrown out, because all `AnalUnit`s
in the file required re-analysis so as to trigger necessary transitive
failures and remove stored compile errors which would no longer make
sense (because a fresh compilation of this code would not emit those
errors, as the units those errors applied to would fail sooner due to
referencing a failed file). Now, this case only applies when a file has
severe top-level errors, which is far less common than something like
having an unused variable.
Lastly, this commit changes a few errors in `AstGen` to become fatal
when they were previously non-fatal and vice versa. If there is still a
reasonable way to continue AstGen and lower to ZIR after an error, it is
non-fatal; otherwise, it is fatal. For instance, `comptime const`, while
redundant syntax, has a clear meaning we can lower; on the other hand,
using an undeclared identifer has no sane lowering, so must trigger a
fatal error.
This fix doesn't matter at all in the grand scheme of things, but I
think the story behind it is perhaps curious, as it might point at a
design flaw in the Sema's error reporting API. So, a story:
On lobsters, there's a rather heated discussion on the merits on RAII vs
defer. I don't really like participating in heating discussions, but
also sort of can't stop thinking about this.
My own personal experience with Zig's defer and errdefer is that they
are fiddly to get right consistency --- if a program has a lot of
resource management to do, I _always_ mess up at least one
defer/errdefer. I've found my internal peace by just avoiding
spread-out, "pox" resource management, and instead centralizing resource
ownership under one of the following patterns:
* Either the thing is acquired and released in main
* Or main allocates N instances of thing, and then the rest of the code
explicitly juggles this finite pool of N. Notably, this juggling
typically doesn't involve defer/errdefer at all, as, at this level of
precision, there are no `try`s left, so you only code the happy path
* Or there's some sort of arena thing, where a bunch of resources have a
single owner, the user's don' bother cleaning up their resources, and
instead the owner does it once at the end.
So I wanted to make a lobster.rs comment in the vein of "yeah, if your
program is mostly about resource management, then Zig could be kinda a
pain, but that's friction tells you something: perhaps your program
shouldn't be about resource management, and instead it should be doing
what it is supposed to do?". And, as an evidence for my claim, I wanted
to point out some large body of Zig code which doesn't have a lot of
errdefers.
So, I cracked opened Sema.zig, `ctrl+f` for `defer`, saw whopping 400
something occupancies, and my heart skipped a bit. Looking at the
occurrences, _some_ of them were non-resource-related usages of defer.
But a lot of them were the following pattern:
```zig
const msg = try sema.errMsg(src, "comptime control flow inside runtime block", .{});
errdefer msg.destroy(sema.gpa);
```
This is exactly the thing that I know _I_ can't get right consistently!
So, at this point, I made a prediction that at least one of `errdefer`s
is missing. So, I looked at the first few `const msg = try` and of
course found one without `errdefer`.
I am at 0.8 that, even with this PR applied, the claim will still stand
--- there will be `errdefer` missing. So it feels like some API
re-design is in order, to make sure individual error messages are not
resources.
Could Sema just own all partially-constructed error messages, and, at a
few known safe-points:
* if the control flow is normal, assert that there are no in-progress
error messages
* if we are throwing an error, go and release messages immediately?
I am unlikely to do the actual refactor here, but I think it's worth
highlighting the overall pattern here.
PS: I am only 0.9 sure that what I've found is indeed a bug! I don't
understand the code, I did a dumb text search, so I _could_ have made a
fool of myself here :P
Previously, stepping from the single statement within the loop would
always exit the loop because all of the code unrolled from the loop is
associated with the same line and treated by the debugger as one line.
